● Private Client, Generative AI Engineer- Built a comprehensive Amazon Bedrock Agent pipeline for creating a Knowledge Base from thousands of PowerPoint, PDF, and Excel reports. Leveraged OpenSearch for the vector database, applied a Foundation Model for parsing the extracted content, and prompted Claude Sonnet 3.5 for generating LLM responses.- Developed a semantic search application with Amazon Bedrock Retrieval and Generation APIs to get relevant content from the S3 report database, including metadata for source document references.● A.Team Client, AI Engineer- Created a scalable agentic system with LangChain and LangGraph on Google Cloud Platform (GCP),deploying a healthcare application with secure API endpoints for the AI Assistant (Anthropic Claude).- Simulated application load testing with Kubernetes and Locust, identifying performance bottlenecks and ensuring the application could handle high traffic volumes, improving Firestore throughput under peak loads.● Grindr, Generative AI Engineer, Chatbot- Developed a Retrieval-Augmented Generation (RAG) pipeline for offline knowledge curation to enhance a GenAI chatbot for the dating application. Leveraged sentence transformers, Postgres vector databases on AWS RDS, and custom prompt sources to supply the LLMs on Amazon Bedrock with additional context. Implemented similarity search and LangChain chat models to retain conversation history.- Created custom evaluators with Patronus AI for grading model output among competing LLMs such as OpenAI, Claude, and Ex-Human, with GitHub Actions and workflows to manage product releases and regression testing. Integrated PortKey's observability suite and AI gateway on both Amazon Bedrock and other custom LLMs.- Leveraged DevOps tools such as Helm, Argo, GitHub Workflows, Docker, and Kubernetes to deploy FastAPI microservices on Amazon EKS clusters. The microservices are written in Kotlin.

● Large Language Models (LLM) for Semantic Search and Substitution, applying sentence transformers, FAISS, and embeddings. T5 implementation with Seldon MLOps. ● Developed Learning-To-Rank (LTR) models for personalization with XGBoost Ranker, improving our customer recommendations.● Created an NLP Evaluation Toolkit for search and substitution by applying fuzzy matching techniques. Derived ground truth datasets from Google Analytics and Elasticsearch.

Streamlining Spark pipelines for concept extraction and patient history classification from semi-structured clinical notes.

● To establish the Data Science practice at Deloitte, I defined the Business Requirements for an AI-driven retail shopping platform, with eleven separate models for causal inference, propensity scoring, affinity analysis, and product recommenders. These requirements were instrumental in securing several years of project funding.● Generated complex, synthetic retailer datasets with Python SDV using a Gaussian Copula model. The tabular data was resampled to capture multivariate relationships among the shopping data model entities.● Created affinity and behavioral scoring models on Azure Databricks using the synthetic data to feed the gradient boosting algorithm, calculating Shapley values to interpret the model output and provide visual explanations of feature importances. Wrote extensive model validation notebooks that tested the model ensembles.

● Persuaded the executive team to sell Synthetic Data, spurring $75M in new contract deals. Implemented algorithms to synthesize anonymized and generalized features to prevent re-identification, calculating both k-anonymity and l-diversity.● Eliminated manual record matching with automated PySpark record linkage techniques for name, address, and merchant matching. The pipeline used tokenizers, N-grams, hashing transformers, and Locality Sensitive Hashing (LSH), a high-dimensional nearest neighbor search.● Generated demographic predictions with a Spark random forest classifier from consumer transactional features. Extracted distribution features with vector assemblers, bucketizers, andother user-defined functions (UDFs).

● Spearheaded a “small but mighty team” to reduce unplanned absences at six North American auto manufacturing plants, innovating weather and event features with NLP (spaCy and NLTK). Combined challenger models (SARIMAX, XGBoost) at multiple levels of aggregation (crew and production line level), including an auto-encoder for anomaly detection to identify outliers with dynamic thresholds.● Streamlined model and data pipelines with Palantir Foundry, a big data platform with continuous integration. The software streamed vehicle sensor data (SQDF, Witech, Vstat, and Data Logger), which was then compressed with Dynamic Time Warping to highlight potential engine or power train problems. The PySpark pipeline chained LSTMs and Chi-Square analysis to identify unique features for out-of-sample cohorts and to predict warranty repairs as well.● Presented the results of a production loss model to the CTO, a non-parametric Monte CarloSimulation and a parametric negative binomial distribution (R fitdistrplus). This simulationestimated lost production units at manufacturing plants, in contrast to a traditional ARIMA timeseries model.● Led our Data Science team with presentations on Shapley Additive Explanations for modelinterpretation; Long Short Term Memory (LSTM) Networks; Apache Spark; Model ProductionPipelines; and State-Space Time Series.

● Led a large team of onshore and offshore developers to deliver over a dozen models to Sam’s Club on their Hadoop platform: Demand forecasting; customer segmentation and clustering (K-means); propensity scoring (multinomial); churn, renewal, and attrition (libsvm); association and basket analysis (R arules); seasonality (X-13 ARIMA); and offer assignment (dynamic bidding algorithm).● Executed biweekly sprints on three-month epics, with customer presentations, demonstrations,and retrospectives on a monthly basis. Consulted with Sam’s Club directors for strategic planningto govern enterprise models with distributed logging, monitoring, and a central repository.● Created a greedy algorithm in Apache Spark as an alternative to a single-threaded auctionalgorithm, achieving a 50X improvement in performance. The algorithm is a traditionaloptimization problem for distributing a fixed number of offers to millions of customers with multipleconstraints. PySpark and RDD API.

● Synthesized automated trading systems from machine learning models to generate predictionswith a proprietary formula language for extracting technical analysis features.● Backtested systems in R and Python, allocating assets using the Kelly Criterion and Optimal f.The R trading package SPLATR was the precursor to AlphaPy.● Experimented with algorithms for statistical arbitrage (pairs trading), optimal portfolio rebalancing,and volatility hedging, for example, Shannon’s Demon. Analyzed runs and sequences to measureserial dependence in pricing time series, for example, the significance of streaks.